docs: streaming LLM responses implementation plan (issue #71)

[TumCucTom]

Summary

Implementation plan for issue #71 — streaming output for LLM responses.

What This Is

This is a design document only — no code changes. The PR contains docs/STREAMING_PLAN.md which covers:

1. API design — new StreamChunk schema, generate_stream() abstract method
2. Implementation — OpenAI and Anthropic streaming clients, partial tool call buffering
3. Agent loop changes — streaming-aware run(), cancellation support
4. Rendering design — thinking vs content rendering rules, ANSI terminal output
5. Compatibility — streaming is opt-out via --no-stream flag

Key Design Decision

Streaming is the default UX. Non-streaming is available via --no-stream for scripted/power-user cases.

Test Plan

• Review plan for correctness and completeness
• Approve so implementation can proceed on a feature branch

Related

• Issue Any plans to support streaming output for LLM responses? #71: Any plans to support streaming output for LLM responses? #71
• PR feat: add streaming output renderer #81 (streaming for tool output — different concern)

🤖 Generated with Claude Code